Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Chainerによる深層学習(3)
Search
Sponsored
·
SiteGround - Reliable hosting with speed, security, and support you can count on.
→
youichiro
March 08, 2017
Technology
160
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
Chainerによる深層学習(3)
長岡技術科学大学
自然言語処理研究室
B3ゼミ発表(第7回)
youichiro
March 08, 2017
More Decks by youichiro
See All by youichiro
日本語文法誤り訂正における誤り傾向を考慮した擬似誤り生成
youichiro
0
1.6k
分類モデルを用いた日本語学習者の格助詞誤り訂正
youichiro
0
140
Multi-Agent Dual Learning
youichiro
1
210
Automated Essay Scoring with Discourse-Aware Neural Models
youichiro
0
150
Context is Key- Grammatical Error Detection with Contextual Word Representations
youichiro
1
170
勉強勉強会
youichiro
0
110
Confusionset-guided Pointer Networks for Chinese Spelling Check
youichiro
0
220
A Neural Grammatical Error Correction System Built On Better Pre-training and Sequential Transfer Learning
youichiro
0
210
An Empirical Study of Incorporating Pseudo Data into Grammatical Error Correction
youichiro
0
240
Other Decks in Technology
See All in Technology
え?フロントエンドエンジニアの ワイがインフラも!?
puku0x
1
690
研究開発部の紹介 / Sansan R&D Profile
sansan33
PRO
4
24k
AWS ネットワーク構築でハマった(ハマりかけた) 5選とそこから得た教訓
nagisa53
4
250
カートの信頼性を担保するWireMockを使ったe2eテスト
ykagano
0
370
MIRU 2026 チュートリアル
keisuke198619
0
970
ブラウザ研修 2026
recruitengineers
PRO
6
1.1k
Sansan Engineering Unit 紹介資料
sansan33
PRO
1
4.9k
強化学習「理論」入門
enakai00
3
3.6k
猫付きpingコマンドを自作
uyuki234
0
250
【AG-UI × A2UI × MCP Apps】Generative UIをやさしく解説する
nrinetcom
PRO
1
140
認知負荷をGemini で溶かす — GKE 基盤「Orbit」における AI エージェントの実践
sansantech
PRO
1
320
Webアクセシビリティ入門 2026
recruitengineers
PRO
3
630
Featured
See All Featured
Imperfection Machines: The Place of Print at Facebook
scottboms
270
14k
Balancing Empowerment & Direction
lara
6
1.2k
Rebuilding a faster, lazier Slack
samanthasiow
85
9.6k
Product Roadmaps are Hard
iamctodd
55
12k
Noah Learner - AI + Me: how we built a GSC Bulk Export data pipeline
techseoconnect
PRO
0
390
Effective software design: The role of men in debugging patriarchy in IT @ Voxxed Days AMS
baasie
0
480
Design of three-dimensional binary manipulators for pick-and-place task avoiding obstacles (IECON2024)
konakalab
0
540
How to optimise 3,500 product descriptions for ecommerce in one day using ChatGPT
katarinadahlin
PRO
2
3.8k
Optimising Largest Contentful Paint
csswizardry
37
3.9k
The Invisible Side of Design
smashingmag
301
52k
Faster Mobile Websites
deanohume
310
32k
SERP Conf. Vienna - Web Accessibility: Optimizing for Inclusivity and SEO
sarafernandez
2
1.6k
Transcript
Chainerによる深層学習 (3) 平成29年3月9日 長岡技術科学大学 自然言語処理研究室 小川耀一朗
Chainerによる分類問題 1/26
プログラムの全体図 2/26
Irisデータ • 150個のアヤメのデータ • データ: 花びらの長さ、幅、がく片の長さ、幅 • アヤメの種類: setosa(0)、versicolor(1)、virginica(2) 例)
[ 5.0999999 3.5 1.39999998 0.2 ] => 0 [ 7. 3.20000005 4.69999981 1.39999998] => 1 [ 6.30000019 3.29999995 6. 2.5 ] => 2 訓練データ → 奇数番目のデータ75個 テストデータ → 偶数番目のデータ75個 3/26
Irisデータの用意 4/26
Irisデータの用意 Irisデータはscikit-learnに付属しているものを使う scikit-learn: 機械学習ライブラリ 分類や回帰、クラスタリングなどの機能が実装されている 5/26
Irisデータの用意 X : 花のデータ >>>print(X) [[ 5.0999999 3.5 1.39999998 0.2
] [ 4.9000001 3. 1.39999998 0.2 ] … [ 5.9000001 3. 5.0999999 1.79999995]] 6/26
Irisデータの用意 Y : 教師データ >>>print(Y) [0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2] 7/26
Irisデータの用意 Y2 : 教師データ(変形) >>>print(Y2) [[ 1. 0. 0.] [
1. 0. 0.] … [ 0. 0. 1.]] 8/26
訓練データ、テストデータの用意 9/26
訓練データ、テストデータの用意 奇数番目のXデータ → 訓練データ 奇数番目のY2データ → 訓練用の教師データ 10/26
訓練データ、テストデータの用意 偶数番目のXデータ → テストデータ 偶数番目のYデータ → 正解データ 11/26
モデルの定義 12/26
モデルの定義 13 入力:花びらの長さ、幅、がく片の長さ、幅 → 4次元 出力:setosa(0)、versicolor(1)、virginica(2) → 3次元 /26
モデルの定義 14 入力層→中間層:シグモイド関数 中間層→出力層:そのまま 損失関数:二乗誤差 /26
パラメータの学習 15/26
パラメータの学習 16/26
パラメータの学習 17 おまじない /26
パラメータの学習 18 ミニバッチ処理 1回のパラメータ更新にランダムに 取り出した25個の訓練データを使う /26
パラメータの学習 19 5000回繰り返す /26
パラメータの学習 20/26
評価 21/26
評価 22 テストデータをモデルに投入し、予測データを得る テストでは勾配を求める必要はないので Variable変数をvolatile=‘on’にする >>>print(ans) [[ 1.01755786e+00 1.39655769e-02 -2.12547127e-02]
[ 9.83523667e-01 3.55108976e-02 -3.01905852e-02] [ 1.03329992e+00 -1.48231089e-02 -1.54979099e-02] … [ -1.24957561e-01 2.79694885e-01 8.36571217e-01]] /26
評価 23 nrow = 75 ncol = 3 /26
評価 24 予測の最大が正解データと一致したら ok+1 >>>print(“{} {}”.format(ans[30], yans[30])) [ 0.0668037 0.80043787
0.15562642] 1 /26
評価 25/26
発表内容 26 • ChainerによるIrisデータの分類 参考文献 「Chainerによる実践深層学習」第4章 新納 浩幸 著 オーム社
/26